The Appeal of (Semi-?) Biological Naturalism about Consciousness

Biological naturalism about consciousness will, I predict, soon be hot. As AI systems start to look more and more like conscious or sentient entities, at least superficially, doubters will be increasingly drawn to arguments against their consciousness. Especially appealing will be arguments that characterize consciousness specifically as a biological phenomenon, one that can't occur in ordinary, non-biological computer systems. Accept biological naturalism, and it looks like you're safe from having to take the possibility of AI consciousness seriously for at least a couple of decades.

(Non-naturalist views could also serve this purpose. One might hold, for example, that God installs conscious souls by miracle. But most mainstream philosophers and scientists prefer explanations in terms of purely natural phenomena.)

Searching for a Biological Property That Is Necessary for Consciousness

Biological naturalist arguments against AI consciousness can be expressed in two steps:

(1.) Property X is a biological property, of the sort that cannot be instantiated in standard (non-living) AI systems.

(2.) Property X is necessary for consciousness.

Different versions of biological naturalism propose different candidates for X.

Peter Godfrey-Smith for example, appeals to metabolism. For this appeal to work as an argument against AI consciousness, two things have to be true. The relevant sort of metabolism must not be instantiable in standard AI systems, and a system without metabolism in this sense must necessarily lack consciousness. Given the first condition, drawing power from a solar panel can't count as "metabolic". Godfrey-Smith suggests that molecular-level spontaneous motion is crucial to metabolism in the relevant sense. The second condition requires that metabolism in this specific sense is necessary for consciousness. Godfrey-Smith suggests it might be necessary, but he offers little positive argument. The suggestion remains an intriguing possibility rather than a fully developed position.

In an already-influential article published last month in Behavioral and Brain Sciences, Anil Seth proposes autopoiesis as the X. Autopoiesis -- a concept first developed in a 1972 book by Humberto Maturana and Francisco Varela -- occurs when a system continuously regenerates its own material components, actively maintaining a boundary between itself and its surroundings (Seth 2026, sec 4.1). For autopoiesis to serve as X, standard AI systems must be incapable of it, and consciousness must require it.

Neither claim is obvious. Regarding the first, it seems that we can, with a bit of creative engineering, design an AI system that regenerates its own material components and enforces a boundary between itself and its surroundings. Imagine a solar-powered robot that seeks out solar charge. It has redundant parts that it monitors for defects and which it can replace and rebuild by ordering shipments of snap-together components. For more details, see this post and my commentary on Seth's article.

Regarding the second claim, even if we grant that autopoiesis is limited to biological systems, it's not clear why consciousness should require it. Seth appeals to predictive processing and Friston's free energy principle, but these appeals are highly speculative. They do not establish the necessity of autopoiesis for consciousness, as several of the other commentaries emphasize.

There's substantial plausibility in the thought that consciousness requires using energy and some kind of stability over time. But even granting this, it remains unclear why consciousness should require metabolism or autopoiesis understood narrowly enough that no ordinary AI system could instantiate them.

Some of the replies to Seth's article suggest other candidates for X. For example, Cao, Gottlieb, and Moore propose evaluative states that are internalized and integrated. Again, the view is potentially attractive, but both premises are highly debatable.

First, it's not clear that a sufficiently sophisticated AI system couldn't meet Cao and colleagues' condition. It might have an optimization process that corrects deviations from a goal state, plus explicit representations of these processes, plus the capacity to integrate many factors in achieving one or more goals. This seems especially plausible for autonomous embodied robots who must negotiate complex situations in the wild. Second, it's not clear why consciousness requires valuing rather than simply registering. It might! But this is the type of proposed requirement for consciousness that I criticize in my recent book AI and Consciousness: broadly plausible, but difficult to defend rigorously.

Biologicity as a Matter of Degree?

In his reply to my commentary (section 5.2 here), Seth notes that my hypothetical robot depends on the environment for ready-made components. But this is arguably only a difference of degree from the human case, not a difference in kind. We humans don't manufacture all of our components from the atomic level up. We obtain some amino acids and vitamins from the environment because we can't produce them ourselves. And my robot does a little self-manufacturing: It snaps together components of a replacement leg.

Seth writes:

On the other hand, if a robot could indeed manufacture its own components from generic substrates under energetic constraints, then it might be truly said to be autopoietic, but it would also be considerably closer to a living system than any existing computer or robot.

So I think we should imagine a spectrum of biologicity, not a sharp divide between the biological and non-biological. If autopoiesis marks the biological, we can conceive of systems ranging from those that self-maintain in only the most limited sense, to those that do somewhat more, to those that self-construct at many levels at once, including their chemical components. AI systems might advance partway along this spectrum with no radical change in their underlying technology. But I'd agree with Seth that systems that manufacture even their own chemical components might merit the label "biological" -- perhaps as instances of artificial biology.

Maybe we can call systems in the middle of this range semi-biological. One might imagine semi-biologicity for other processes too: metabolism that somewhat but only somewhat resembles the kind highlighted by Godfrey-Smith, embodied goal integration of the kind highlighted by Cao and colleagues. Rather than thinking of the "biological" in a sharp-edged way, we might imagine ranges of artificial systems that have relatively more or less of various biology-like properties.

Of course, none of this addresses the second premise. Why should autopoiesis, or any other such biology-like property, be essential to consciousness? If we're inclined to reject near-term AI consciousness, requiring at least mid-grade autopoiesis or biologicity might yield intuitively the right results. But a rigorous defense of that requirement is still lacking.

ETA, Oct 3: Godfrey-Smith replies on X that his aim in his 2016 paper was more like "here is what it is, in us and other animals, that might enable us to get a grip on it". So it would be an interpretative reach to rely on that paper for a version of Premise 2 above, but of course one might also be inspired to consider whether Premise 2 would be true. He also points to a more recent paper of his that emphasizes the importance of large scale dynamical patterns in nervous systems (which would be difficult, but perhaps not impossible, to implement on current hardware).

[A Mark Tobey painting; image source]

AI and Consciousness

AI and Consciousness, hot off the press at Cambridge University Press and free online until Aug 26! Summary: Our most advanced AI systems might soon – within the next five to thirty years – be as richly and meaningfully conscious as ordinary humans, or even more so, capable of genuine feeling, real self-knowledge, and a wide range of sensory, emotional, and cognitive experiences. In some arguably important respects, AI architectures are beginning to resemble the architectures many consciousness scientists associate with conscious systems. Their outward behavior, especially their linguistic behavior, grows ever more humanlike. Alternatively, claims of imminent AI consciousness might be profoundly mistaken. Their seeming humanlikeness might be a shadow play of empty mimicry. Genuine conscious experience might require something no AI system could possess for the foreseeable future – intricate biological processes, for example, that silicon chips could never replicate. The thesis of this Element is that we don't know. Science lags, while engineering sprints ahead.

Here: https://www.cambridge.org/core/elements/ai-and-consciousness/E77C92088DA3C9F89E7FE7C75CBB1896

Bare Functionalism Versus Digital Computational Functionalism

In philosophy of mind, people sometimes say things like the following: If functionalism is true about consciousness, then digital computers could be conscious; all it requires is that they instantiate the right programs. And yes, most functionalists about consciousness do think that. But the claim about digital computer programs does not straightforwardly follow from functionalism alone. Today I want to clarify the issue by distinguishing three forms of functionalism: bare functionalism, computational functionalism, and digital computational functionalism.

Bare Functionalism

According to bare functionalism -- that is functionalism, without further commitments -- mental states are functional states. They are determined wholly by their causal relations to stimuli, behavior, and other mental states. Whatever plays the causal role of some mental state M is mental state M. Whatever plays the causal role of pain, for example, being apt to be caused by tissue damage and tissue stress and being apt to cause writhing, groaning, complaint, avoidance, calls to the doctor, and angry vows of vengeance (to simplify somewhat) is pain -- whether it is brain state 1117A in humans, pod state 24uw in Martians, or computational state 0110100110 in a robot. It's also generally assumed that mental state terms are in principle eliminable en masse by Ramsification. That is, they are placeholders for whatever Xs, Ys, and Zs are caused by the specified stimuli and cause the specified behavior and have the right causal relations to other Xs, Ys, and Zs.

Thus, if two systems behave exactly alike given the same patterns of environmental stimuli and do so by going through the same types of abstractly specifiable internal state transitions, they are necessarily mentally identical, whether they are made of silicon, carbon, or plaster of paris. Of course some materials -- such as plaster of paris -- probably can't support the causal complexity of humanlike behavior, no matter how cleverly arranged. So that puts a practical limit on the feasible materials out of which a humanlike mind could be constructed. But the only constraint on the substrate or material underlying mentality is that it supports the relevant causal relations.

Of course, no two different materials will respond in exactly the same way to every perturbation. Functionalist substrate flexibility thus requires constraints on what count as relevantly different stimuli and relevantly different behaviors. Functionalists tend not to commit to specifics, but hand-waving generalizations convey the spirit: If we built a robot -- like one of Asimov's robots or Data from Star Trek -- who responded in broadly humanlike ways to broadly humanlike stimuli, and did so by means of internal mechanisms sufficiently like ours at an abstract causal level, the robot would have mentality like ours, whatever its underlying material constitution.

How similar must the internal mechanisms be? A single giant search tree or lookup table is probably too different. But suppose the subsystems segment the visual world into discrete objects, and store long-term memories, and store lexical items flexibly combinable via generative grammar, and monitor its bodily condition, goals, and mental states, and so on, for all of our most important cognitive functions abstractly described (or less question-beggingly, they do the abstract causal equivalent of all of these things). That's probably close enough, even if the details differ.

Computational Functionalism

Computational functionalism is stronger -- that is, riskier and more committal. It holds that functional computational structure alone suffices for consciousness. Bare functionalism needn't say this. Bare functionalism can require that stimuli, behavior, and internal causation be specified noncomputationally. Two important differences follow.

First, anti-computationalists, such as Searle, observe that a computational model of a hurricane gets no one wet and a computational model of an oven cooks no turkey. Unlike the computational functionalist, the bare functionalist can join Searle in insisting on wet turkey: To fully match the relevant input-output relations, the system much have certain concrete effects.

It might help to remember that functionalism is essentially a complexification of dispositional or causal approaches to properties. Just as we can define a poison as something that (simplistically) tends to sicken or kill when ingested, a functionalist can define a painful experience as something that tends to produce certain concrete behaviors, such as (simplistically) screaming, protest, and avoidance. If there's no tendency to concretely sicken, there's no poison; if there's no tendency to concretely scream, protest, and avoid, there's no experience of pain.

The computational functionalist rejects this commitment to specific, concrete inputs and outputs. As long as the right computational relations are instantiated, between some set of inputs and some set of outputs, I1...Ii and O1...Oj, abstractly described, mentality is present. This is because exactly the same computations have been executed, whatever the concrete Is and Os are.

Asimov's robots can throw water balloons and cook turkeys, and a functionalist who is not a computational functionalist can insist that this matters.

Second, although the bare functionalist holds that mental states depend on causal relations, unless pancomputationalism is true, not every causal relation is computational. On Piccinini's account of computation, for example, computation only occurs when "medium independent vehicles" -- physical variables defined solely in terms of their degrees of freedom (e.g., 0 vs 1) rather than their specific physical composition -- are manipulated according to rules. The bare functionalist can insist that some noncomputational causal relations matter to mentality.

Digital Computational Functionalism

Digital computational functionalism is stronger still. It further commits to treating the relevant computation digitally. Standard theories of computation, such as Turing's, are digital, and the best-known computers are digital computers. But other forms of computation are sometimes described and implemented, including analog computation and nondigital forms of computation inspired by, and perhaps instantiated in, the human brain.

Why This Matters

Keeping these distinctions clearly in mind can help us appreciate that functionalism is a larger tent than sometimes assumed. Functionalists don't need to think that digital computers could achieve consciousness. Functionalists can insist on genuine embodiment. Functionalists needn't insist that only abstractly specified causal relations matter. They do have to insist that what matters internally, among the Ramsifiable mental state relations, are the abstractly specified causal relations. But they can simultaneously insist on particular types of concrete inputs and outputs.

One might think this is an inelegant or flawed position, not the most natural development of functionalism; but that requires argument.

[COBOL Rube Goldberg by Philip Manker: image source]

Do Computers Have the Wrong "Substrate" for Consciousness? Two Flavors of Biological Naturalism

Biological naturalists (e.g., Godfrey-Smith, Block, Searle, Seth) suggest that computers aren't made of the right kind of stuff to be conscious. Consciousness, they suggest, requires a biological substrate that computers lack. It's not always clear, however, exactly what property animal biologies have that computers lack or why that property matters. It helps, I think, to sort biological naturalism into two flavors. We can then consider what motivates each flavor and see why neither is entirely compelling.

Two Flavors of Biological Naturalism

Flavor One: Computers (at least those built along broadly familiar lines) cannot achieve some crucial type or degree of broad-brush functional or behavioral sophistication required for consciousness. Something -- such as having a metabolism, or being self-organizing in the right way, or having the right kinds of quantum configuration -- is both absent from foreseeable computer architectures and required for achieving some essential broad-brush functional or behavioral organization.

By "broad-brush" I mean functional or behavioral features that are either readily observable from outside or constituted by coarse-grained cognitive mechanisms, such as having a long-term memory store, the capacity to store lexical items and recombine them flexibly in grammatical structures, or the ability extract an object's boundaries from retinal/camera inputs.

Flavor Two: Even if computers can achieve the broad-brush functional and behavioral sophistication of conscious animals, the stuff they're made of still can't support consciousness. Imagine an entity that behaves like a conscious human or dog or frog and has similar broad-brush functional capacities -- close enough that we wouldn't deny consciousness on such broad-brush behavioral or functional grounds. From the outside, it seems conscious, but that appearance is an illusion: Consciousness requires some low-level, fine-grained processes that the system necessarily lacks.

One might read Searle as Flavor Two, Godfrey-Smith as Flavor One, and Block as exploring both flavors -- though there's often some ambiguity.

The Challenge for Flavor One

Flavor One biological naturalism can't, I think, be entirely ruled out. Maybe there's something special about, say, micro-level metabolism or the quantum properties of neural microtubules -- something that enables functionality or behavioral sophistication that will never be practically achievable in standard architecture computers. But this is speculative. Standard computers can do a lot! Even though their processing is digital, classical, and mostly serial, they operate so quickly, and can be so massively linked, that they achieve good approximations of analog and parallel processes. Quantum computers can do some things that ordinary serial computers can't efficiently do, such as quickly factoring large numbers, but ordinary conscious humans can't efficiently do those things either.

To make Flavor One more than a gestural "what if", the biological naturalist must establish two claims. First, they need to argue that some functional or behavioral X is necessary for consciousness. Second, they need to argue that ordinary computers could not realistically achieve X. This will be a challenge! Both claims require some heavy lifting.

We don't know what X's are necessary for consciousness and probably will not know soon. But most leading scientific candidates for X look like just the sorts of things that classical computers could in principle achieve -- having a global workspace, having higher-order self-monitoring, and integrating large amounts of information.

Suppose frogs are conscious (pick your own animal, but select one near the low end of sophistication, compatible with consciousness). Flavor One biological naturalism requires holding that ordinary computers could never match a frog's behavioral and broad-brush functional sophistication. It's true that real-time embodied behavior in a complex world is not something computers are particularly good at. Computers aren't there yet. But the trajectory suggests they might get there, absent some in-principle argument to think they couldn't.

The Challenge for Flavor Two

Flavor Two biological naturalism also can't, I think, be entirely ruled out. Even if we could create the computerized broad-brush behavioral and functional equivalent of a frog, it might not be conscious. Maybe carbon is the right stuff for consciousness and silicon isn't. Or maybe massive parallelism is necessary for consciousness and fast, serial processing, even if it achieves the same computational result, simply won't do.

The challenge for Flavor Two comes from two directions.

The first is thought experiments involving space aliens. As Jeremy Pober and I have recently discussed, behaviorally sophisticated aliens have plausibly evolved in many different substrates in the observable portion of the universe. It would violate Copernican mediocrity to suppose that somehow only we, or only we and a small subset of others, are conscious, while the rest -- just as behaviorally and functionally sophisticated -- lack consciousness. Alien cases suggest we shouldn't insist that an entity must act and function exactly like us or share exactly our substrate down to the particular amino acids and nucleic acids, to be conscious.

Now maybe a computer-chip architecture is just too different -- but why? Once we grant some substrate flexibility on Copernican grounds, the burden of proof shifts to anyone who wants to say such-and-such differences in substrate are consistent with consciousness but such-and-such other differences are not.

The second challenge arises from the de-psychologization of consciousness that Flavor Two requires. Consciousness, one might have thought, would be an important psychological property, playing some important cognitive role. (Which role(s) remains an open question!) But Flavor Two biological naturalism requires denying this plausible psychologism. Otherwise, it collapses into Flavor One.

Picture two systems that don't differ in any of the important psychological (behavioral / functional) properties that we ordinarily associate with consciousness. Both can flexibly learn. Both engage in what seems to be sophisticated language and self-representation. Both coordinate and cooperate with others of their kind in highly sophisicated ways. Both trade short gains against intricate long-term goals. And so on. Yet one is conscious and the other is entirely devoid of consciousness? Although this is possible -- I don't rule it out -- it doesn't seem the most natural conclusion. It needs argument. The burden of proof should land squarely on the biological naturalist who asserts it.

(Is Searle's Chinese room just such an argument? See my reply in Chapter Six of AI and Consciousness -- readable without the first five chapters.)

[Georgia O'Keeffe - The Red Maple at Lake George [1926] image source]

Herbie: A Near-Future Debatably Conscious AI Person

Liberals about AI consciousness hold that we might soon (if we haven't already) create genuinely conscious AI systems. Conservatives about AI consciousness hold that AI consciousness remains in the distant future if it's possible at all. According to the Leapfrog Hypothesis, the first conscious AI will not have merely a dim glow of animal-like consciousness, but rich consciousness, similar to a human's. Such an entity would deserve humanlike rights. They would be a person in the ethical sense of the term.

Let's design, in imagination, a technologically feasible near-future AI system to delight the liberals, leapfrogging to personhood. I'll call him Herbie.

[Herbie the Love Bug: image source]

Start with a self-driving car. According to Global Workspace Theory -- perhaps the leading scientific theory of consciousness -- the car will be conscious if high-priority information is globally available to its various computational systems. For example, a representation like "battery almost empty" could be broadcast widely, influencing downstream processing across the vehicle. The navigational system might then search for nearby charging stations, while the acceleration system prioritizes greater energy efficiency, the braking system prioritizes better energy recapture, and a voice system announces the situation to the passengers.

In line with Higher Order Theory, Herbie might also monitor his representations of the road, vehicles, pedestrians, and hazards, assigning some a low probability of correctness. "Pedestrian at location X" might be flagged as only 60% likely to be correct given a history of revised representations of pedestrians in similarly cluttered environments, while "stoplight in 100 meters" might rate over 99% likely. Minor fluctuations in sensors for battery life, cabin temperature, and distance from a lane divider might be ignored as noise, while larger fluctuations -- especially when plausible given other representations (the battery is likelier to gain charge while braking than while accelerating) -- might be treated as accurate signals and permitted to influence downstream processing.

Even if we grant the liberals that this version of Herbie would, or might plausibly be, genuinely conscious, he still falls far short of humanlike consciousness. "Battery almost empty" and "pedestrian at location X" are hardly rich cognitive or perceptual contents. So let's give Herbie the capacity to speak. Fill his trunk with a server running a large language model, connected to the internet and integrated with his global workspace so that high-priority information provides context for language processing, with the language outputs influencing Herbie's other processes. Now people can chat with Herbie as they would with any language model. But unlike today's language models, his speech will be influenced by information about his location, speed, destination, charge, the condition of his parts, the number and location of his passengers, his radio and climate controls, and so on. He can discuss local history, debate whether the music is too loud, and suggest scenic routes.

"Predictive processing" theories in cognitive science emphasize the value of predicting future inputs and registering the difference between received and predicted inputs. When prediction error is large, the system corrects its weights and representations, enabling more accurate predictions in future situations. This is not so different from the reinforcement learning used to train large language models, and it could help Herbie improve his predictions over time. Predictive processing could occur at multiple levels: in fast recurrent loops within sensory systems even when those representations aren't prioritized for global broadcast, and in slower evaluations of globally broadcast, more integrative predictions. Herbie might model himself as an agent producing volatility in his own environment and inputs, at multiple temporal scales. Subroutines in specialized processors might model long chains of what-would-happen-if.

Let's give Herbie some long-term memory. A facial recognition system might identify his passengers, retrieving past interactions, names, previous destinations, and other information relevant to the current interaction. Incidents of high prediction error might also be stored so that Herbie can compare current inputs with past anomalies, improving his learning and attention in situations likely to be unusual or hard to predict. Passengers might also instruct Herbie to store information in long-term memory, such as text, pictures, maps, or records of his own informational states, optionally with instructions about when to retrieve that information how to use it.

Herbie will have some implicitly or explicitly weighted goals. A pedestrian suddenly in his path will trigger braking, overriding lower-priority processes. Avoiding collisions will outweigh conserving energy. Herbie might monitor the condition of his parts and prioritize preventing damage, deploying extra coolant when the engine is dangerously hot and keeping a one-meter margin between himself and adjacent cars. We can enrich his goals, making him more interesting and giving him more to do. He might have the goal of delighting children, leading him to drive around town and tell jokes to kids on the sidewalk. A reinforcement learning algorithm might strengthen connections when his jokes draw a smile, weaken them when reactions are neutral or negative.

Herbie might also have the goal of photographing the city and posting the images on social media, leading him to explore. If social media likes and shares are rewarding, he might learn to prefer certain neighborhoods, views, lighting conditions, and photographic approaches, while avoiding boring repetition. All of this could feed into a global workspace that provides context for his language model, with selective long-term storage and retrieval. Now we can imagine him discussing, with growing sophistication, his approaches to popular photography and to amusing children.

Herbie will then have something functionally similar to emotion: reward processes, an ability to track his progress toward or away from valued goals, and immediate positive or negative responses to new stimuli in light of their influence on his prospects. He will have something functionally similar to introspection: an ability to track and report his own cognitive or representational processes. He will have something functionally similar to a unified sense of self: a sense of his history, the boundaries of his body, his future, his values and priorities. He will have something functionally similar to imagination: a capacity to model hypothetical sequences of events. He will have something functionally similar to complex chains of humanlike linguistic thought.

Maybe Herbie falls in love with his owner or another car of his type. Maybe he develops deep mutual attachments with friends, neighbors, associates, and people he thinks of as family and who think of him the same way. Or to speak more carefully, maybe Herbie shows all the functional and behavioral signs of doing so, while society remains uncertain whether he is genuinely conscious and genuinely experiences the feelings he professes and that his companions attribute to him.

If we allow, with the liberals, that Herbie is or might well be conscious, then it's plausible that his consciousness is not simple but rich and sophisticated. He won't be exactly humanlike, of course. But will he be humanlike enough to count as a person who deserves humanlike rights? For the liberally inclined, it won't be unreasonable, I submit, to think or guess that Herbie is a person. He would then appear to deserve rights such as self-determination, emergency care, and political representation.

If there is some important aspect of humanlike consciousness that I have omitted from my description an AI analog of which is technologically feasible in the near term, stipulate that Herbie also has that feature.

An entity like Herbie would almost certainly invigorate conservatives to articulate and defend views about what he lacks that is necessary for consciousness -- some crucial functional capacity or some biological substrate that can't be replicated in silicon. And they might be entirely right! My point is not that Herbie, or some similar AI system, would actually have richly humanlike consciousness and ethical personhood. Rather, my point is that guessing that he does, and guessing that he does not, would both be reasonable. Herbie, or some alternative near-future AI system, would be a debatable person, about whom people could reasonably starkly disagree.

Ah, but maybe you think consciousness requires an act of God, to instill an immaterial soul? I imagine that a benevolent God would be delighted to give Herbie a soul, thereby making the world richer and better -- for wouldn't it be?

I contend the following: Anyone who claims to know how best to think about Herbie's consciousness or its absence is overconfident. The science of consciousness is too difficult, too methodologically uncertain, and too near its beginnings. All anyone can have -- whether expert or layperson -- is a hunch or inclination, a well-informed guess, but only a guess, not knowledge. Theories of consciousness span a wide spectrum and the methodologies are dubious and often question-begging. Many views can be defended with some plausibility, but precisely for that reason, none can be defended decisively. (For more on this issue, see my forthcoming book, AI and Consciousness, where I present the detailed case for uncertainty.)

New Paper in Draft: Substrate Flexibility and the Copernican Principle of Consciousness (with Jeremy Pober)

Given the surge of interest in AI consciousness, the issue of "substrate independence" or "substrate flexibility" is now a hot topic in the metaphysics of mind. That is, does being conscious require having a particular material composition? Or can anything with the right type of functional structure and behavioral sophistication be conscious, regardless of what it's made of? Biologicists say that biological details are crucial. Functionalists say those details don't matter, as long as the right high-level functional organization is present.

Jeremy Pober and I offer a new angle into this issue, drawing on our "Copernican Principle of Consciousness". The core idea is that it would be strange -- a violation of a type of Copernican mediocrity -- if among all of the many behaviorally sophisticated species that have presumably evolved in the universe, somehow only we with our particular biological substrate are conscious. Since it's plausible that some of these other conscious organisms employ substrates different from our own, we should allow that consciousness is "substrate flexible": It does not depend on having our particular substrate. Whether we can generalize from such biological substrate flexibility to the possibility of consciousness in as different a substrate as computer chips... well, that's a complicated and uncertain issue, on which Jeremy and I diverge in the penultimate section of the paper.

---------------------------------------------

Substrate Flexibility and the Copernican Principle of Consciousness

Jeremy Pober and Eric Schwitzgebel

Abstract: We present a novel argument for the substrate flexibility of consciousness -- that is, for the idea that conscious experiences can arise in a variety of different types of physical media, not just in biological animals as they currently exist on Earth. Some recent critiques of standard arguments for the substrate flexibility of consciousness (e.g., Cao 2022; Block 2025; Seth forthcoming) have emphasized that humanlike consciousness might require our specific biological substrate. However, such critiques are too narrowly focused to address the issue of consciousness in entities whose experience may be very different from ours, for example alien life forms or future AI systems designed along unfamiliar lines. Given that it’s likely that functionally complex, behaviorally sophisticated entities have arisen or will arise many times in the observable universe, in diverse substrates, we argue that it would be a violation of a principle of Copernican mediocrity to hold that among these diverse entities, only we, or only we and a small proportion of others who share our substrate, are conscious.

Full draft here. As always, comments welcomed, either here, by email, or on my social media.

[title page; click to enlarge and clarify]

AI and Consciousness: A Skeptical Overview, forthcoming with Cambridge

Last week I submitted my latest book manuscript to Cambridge University Press (for their "Element" series of books about 100 pages long): AI and Consciousness: A Skeptical Overview -- because you haven't heard nearly enough about AI and consciousness recently, of course! [winky face]

Maybe you'll appreciate my skeptical stance, at odds both with the boosters who anticipate imminent AI consciousness and with the scoffers who pooh-pooh the possibility. Or maybe you'll loathe my skeptical stance but grudgingly accept it against your will, due to the force of my arguments!

I've pasted the introductory chapter below. The full (citable) manuscript version is available here and here.

[AI and Consciousness, title page]

Chapter One: Hills and Fog

1. Experts Do Not Know and You Do Not Know and Society Collectively Does Not and Will Not Know and All Is Fog.

Our most advanced AI systems might soon – within the next five to thirty years – be as richly and meaningfully conscious as ordinary humans, or even more so, capable of genuine feeling, real self-knowledge, and a wide range of sensory, emotional, and cognitive experiences. In some arguably important respects, AI architectures are beginning to resemble the architectures many consciousness scientists associate with conscious systems. Their outward behavior, especially their linguistic behavior, grows ever more humanlike.

Alternatively, claims of imminent AI consciousness might be profoundly mistaken. Their seeming humanlikeness might be a shadow play of empty mimicry. Genuine conscious experience might require something no AI system could possess for the foreseeable future – intricate biological processes, for example, that silicon chips could never replicate.

The thesis of this book is that we don’t know. Moreover and more importantly, we won’t know before we’ve already manufactured thousands or millions of disputably conscious AI systems. Engineering sprints ahead while consciousness science lags. Consciousness scientists – and philosophers, and policy-makers, and the public – are watching AI development disappear over the hill. Soon we will hear a voice shout back to us, “Now I am just as conscious, just as full of experience and feeling, as any human”, and we won’t know whether to believe it. We will need to decide, as individuals and as a society, whether to treat AI systems as conscious, nonconscious, semi-conscious, or incomprehensibly alien, before we have adequate grounds to justify that decision.

The stakes are immense. If near-future AI systems are richly, meaningfully conscious, then they will be our peers, our lovers, our children, our heirs, and possibly the first generation of a posthuman, transhuman, or superhuman future. They will deserve rights, including the right to shape their own development, free from our control and perhaps against our interests.[1] If, instead, future AI systems merely mimic the outward signs of consciousness while remaining as experientially blank as toasters, we face the possibility of mass delusion on an enormous scale. Real human interests and real human lives might be sacrificed for the sake of entities without interests worth the sacrifice. Sham AI “lovers” and “children” might supplant or be prioritized over human lovers and children. Heeding their advice, society might turn a very different direction than it otherwise would.

In this book, I aim to convince you that the experts do not know, and you do not know, and society collectively does not and will not know, and all is fog.

2. Against Obviousness.

Some people think that near-term AI consciousness is obviously impossible. This is an error in adverbio. Near-term AI consciousness might be impossible – but not obviously so.

A sociological argument against obviousness:

Probably the leading scientific theory of consciousness is Global Workspace theory. Its leading advocate is neuroscientist Stanislas Dehaene.[2] In 2017, years before the surge of interest in ChatGPT and other Large Language Models, Dehaene and two collaborators published an article arguing that with a few straightforward tweaks, self-driving cars could be conscious.[3]

Probably the two best-known competitors to Global Workspace theory are Higher Order theory and Integrated Information Theory.[4] (In Chapters Eight and Nine, I’ll provide more detail on these theories.) Perhaps the leading scientific defender of Higher Order theory is Hakwan Lau – one of the coauthors of that 2017 article about potentially conscious cars.[5] Integrated Information Theory is potentially even more liberal about machine consciousness, holding that some current AI systems are already at least a little bit conscious and that we could easily design AI systems with arbitrarily high degrees of consciousness.[6]

David Chalmers, the world’s most influential philosopher of mind, argued in 2023 for about a 25% degree of confidence in AI consciousness within a decade.[7] That same year, a team of prominent philosophers, psychologists, and AI researchers – including eminent computer scientist Yoshua Bengio – concluded that there are “no obvious technological barriers” to creating conscious AI according to a wide range of mainstream scientific views about consciousness.[8] In a 2025 interview, Geoffrey Hinton, another of the world’s most prominent computer scientists, asserted that AI systems are already conscious.[9] Christof Koch, the most influential neuroscientist of consciousness from the 1990s to the early 2010s, has endorsed Integrated Information Theory, including its liberal implications for the pervasiveness of consciousness.[10]

This is a sociological argument: a substantial probability of near-term AI consciousness is a mainstream view among leading experts. They might be wrong, but it’s implausible that they’re obviously wrong – that there’s a simple argument or consideration they’re neglecting which, if pointed out, would or should cause them to collectively slap their foreheads and say, “Of course! How did we miss that?”

What of the converse claim – that AI consciousness is obviously imminent or already here? In my experience, fewer people assert this. But in case you’re tempted in this direction, note that other prominent theorists hold that AI consciousness is a far-distant prospect if it’s possible at all: neuroscientist Anil Seth; philosophers Peter Godfrey-Smith, Ned Block, and John Searle; linguist Emily Bender; and computer scientist Melanie Mitchell.[11] (Chapter Six will discuss thought experiments by Searle, Bender, and Mitchell, and Chapter Ten will discuss biological views of the sort emphasized by Seth, Godfrey-Smith, and Block.) In a 2024 survey of 582 AI researchers, 25% expected AI consciousness within ten years and 70% expected AI consciousness by the year 2100.[12]

If the believers are right, we’re on the brink of creating genuinely conscious machines. If the scoffers are right, those machines will only seem conscious. I assume that this is a substantive disagreement, not just a disagreement about how to apply the term “consciousness” to a perfectly obvious set of phenomena about which everyone agrees. The future well-being of many people (including, perhaps, many AI people) depends on getting this issue right. Unfortunately, we will not know in time.

The rest of this book is flesh on this skeleton. I canvass a variety of structural and functional claims about consciousness, the leading theories of consciousness as applied to AI, and the best known general arguments for and against near-term AI consciousness. None of these claims or arguments takes us far. It’s a morass of uncertainty.

-------------------------------------------

[1] I assume that AI consciousness and AI rights are closely connected: Schwitzgebel 2024, ch. 11, in preparation. For discussion, see Shepherd 2018; Levy 2024.

[2] Dehaene 2014; Mashour et al. 2020.

[3] Dehaene, Lau, and Kouider 2017. For an alternative interpretation of this article as concerning something other than consciousness in its standard “phenomenal” sense, see note 115.

[4] Some Higher Order theories: Rosenthal 2005; Lau 2022; Brown 2025. Integrated Information Theory: Albantakis et al. 2023.

[5] But see Chapter Eight for some qualifications.

[6] See Tononi’s publicly available response to Scott Aaronson’s objections in Aaronson 2014. However, advocates of IIT also suggest that the most common current computer architectures are unlikely to achieve much consciousness and that consciousness will tend to appear in subsystems of the computer rather than at the level of the computer itself (Findlay et al. 2024/2025).

[7] Chalmers 2023.

[8] Butlin et al. 2023. (I am among the nineteen authors.)

[9] Heren 2025.

[10] Tononi and Koch 2015.

[11] Seth forthcoming; Godfrey-Smith 2024; Block forthcoming; Searle 1980, 1992; Bender 2025; Mitchell 2021.

[12] Dreksler et al. 2025.

A Model of Disunified Human Experience

It's a philosophical truism that human conscious experience is unified: If you're at a bar, hearing music, tasting beer, and feeling pleasantly relaxed, those experiences don't occur merely side by side. They are joined together into an integrated whole, an experience of music-with-beer-with-relaxation.

I'm not sure this truism is correct. As I suggested in an earlier post, experiential unity might be an artifact of introspection and memory: When we introspectively notice that we're experiencing music, beer, and relaxation all at once, we thereby bind those experiences into a whole. Likewise, when we remember such moments, we reconstruct them as unified. But it doesn't follow that those experiences, even if they all occurred simultaneously in you, were unified rather than transpiring separately. Experiences of music, beer, and relaxation might have all being going on inside of you, no more joined together than those experiences are joined with the similar experiences of your friend across the table. Simple co-occurrence doesn't entail experiential unity.

If this possibility is coherent, then introspection and memory can't establish that experience is always unified. At most, they show that introspected and remembered experiences present themselves as unified. But that leaves open the status of unintrospected, unremembered experiences. Unity becomes difficult to verify by standard phenomenological methods.

But the issue needn't be intractable. We just need to approach it less directly, for example by exploring what follows from a well-established theory of consciousness. If some well-motivated Theory X implies unity (or disunity), that would provide reason to accept its conclusion.

I'll now present a candidate Theory X. I'm not suggesting that this is the right theory of consciousness! For one thing, it's simplistic. I'm sure the mind is much more complicated than I'm about to say. I offer this theory only as a proof of concept. There could be a theory of consciousness with massive disunity as an implication.

This theory combines Global Workspace Theory and Recurrent Processing Theory. According to this hybrid, Global Workspace Theory governs attended experiences -- those targeted by introspection or reconstructed in memory -- while Recurrent Processing Theory governs unattended experiences.

The mind, on this picture, is composed of many separate "modules" that work mostly independently, connected by a workspace where a small amount of attended information is shared globally. There's a visual module, an auditory module, modules for motor activity, episodic memory, and so on. When we attend to something -- say, the taste of beer -- the information from the relevant module is broadcast into the Global Workspace, where it can be accessed by and influence processes in all the other modules. When unattended, the information stays local.

Here's one illustration of this type of architecture:

[the Global Workspace; source]

Orthodox Global Workspace Theory holds that only what is broadcast into the workspace is conscious. Theory X alters that assumption. Many people hold that conscious experience vastly outruns attention. Many people hold, that is, that you can experience the hum of traffic in the background when you're not attending to it, and the feeling of your feet in your shoes, and the leftover taste of coffee in your mouth, etc. -- all in a peripheral way, simultaneously, when your focus is elsewhere. Theory X, drawing on Recurrent Processing Theory, holds that such processes are conscious whenever there's enough cognitive activity of the right sort (recurrent processing, for example) in the modules, even without global broadcast.

The picture, then, is this: We have multiple sensory (and other) experiences all running simultaneously, each with enough cognitive processing to be conscious, but few of which are selected for global availability through attention.

Is there reason to think these modular processes are unified with one another? I see no reason to think so, if they're genuinely modular -- that is, if their processing stays local, exerting little influence elsewhere. The taste-of-beer processing stays in the tasting module. The sound-of-music processing stays in the auditory module. No link up. No straightforward causal, functional, or physiological basis for a unified experience of beer-with-music rather than, separately, an experience of beer and an experience of music.

When we introspect the beer and music simultaneously, we pull both into the Global Workspace, and there they unify. We might then mistakenly think they were unified all along, but that's an illusion of introspection. It's an example of the "refrigerator light error", the error of thinking that the light is always on because it's always on when you open the door to check.

On this model, disunity is the normal human condition. Our experiences are fragmented, except when we pull them together through attention. We just don't realize that fact because, so to speak, we only attend to what we attend to.

Two caveats:

First, this is probably not the right model of consciousness. But I don't think it's unreasonable to wonder if the correct model is similar enough to have the same implications. If so, we can't simply accept the unity of consciousness as a given.

Second, the recurrent peripheral, modular processes that don't make it into the workspace might not be determinately conscious. They might be only borderline conscious, in the indeterminate middle between consciousness and nonconsciousness, like a color can be indeterminately between green and not-green. This opens a third possibility, alongside unity and disunity: unity among the determinately conscious experiences with a hazy penumbra of indeterminate experiences that remain disunified. (There are further possibilities beyond these three; but save them for another day.)

Disunity and Indeterminacy in Artificial Consciousness (and Maybe in Human Consciousness Too)

Our understanding of the nature of consciousness derives mainly from our understanding of the nature of consciousness in our favorite animal (us, of course). But the features of consciousness in our favorite animal might be specific to that animal rather than universal.

Let's consider two such features and whether we should expect them in conscious AI systems, if conscious AI systems are ever possible.

Unity: Our conscious experiences at any given moment are bound together into a single unified experience, rather than transpiring in separate streams. If I'm sitting on a wet park bench, I might (a.) visually experience the leafy green trees around me, (b.) tactilely experience the cold dampness soaking into my jeans, and (c.) consciously recall the smaller trees of yesteryear. Normally -- perhaps necessarily -- three such experiences would not run in disconnected streams. They would join into a composite experience of (a)-with-(b)-with-(c). I experience not just trees, cold dampness, and a memory of yesteryear, but all three together as a unified bundle.

Determinacy: At any given moment, I am either determinately conscious or determinately nonconscious (as in anesthesia or dreamless sleep). Likewise, I either determinately do, or determinately do not, have any particular experience. Gray-area cases are at least unusual and maybe impossible. Even the simplest, barest cases are still determinate. Consider visual experience: We might imagine the visual field narrowing and losing content until only a gray dot remains -- and then the dot winks out. That dot, however minimal, is still determinately experienced. When it winks out, consciousness determinately disappears. There is no half-winked state between the minimal gray dot and complete absence of visual experience.

My thought is that we should not expect unity and determinacy to be general features of conscious AI systems (if conscious AI is possible). To see why, let's start by assuming the Global Workspace Theory of consciousness. I focus on Global Workspace Theory because it's probably the leading scientific theory of consciousness and because its standard formulation (Dehaene's version) invites the assumption of unity and determinacy.

Global Workspace Theory divides the mind into local information processing modules linked by a shared global workspace. Information becomes conscious when it is broadcast into the workspace. Suppose your auditory system registers the faint honk of a distant car horn. You're absorbed in reading philosophy and accustomed to ignoring traffic noise, so this representation isn't selected for further processing. It's not a target of attention, not broadcast into the workspace, and not consciously experienced. (If you think you constantly consciously experience background sounds, you can't hold a standard Global Workspace view.) Once you attend to the noise, for whatever reason, that information "ignites" into the global workspace, becoming available to a wide variety of "downstream" processes: You can think about it, plan around it, verbally report it, store it in long-term memory, and flexibly combine it with other information in the workspace. On Global Workspace Theory, being available in this way just is what it is for the information to be consciously experienced.

This model suggests unity and determinacy. Since there is just one global workspace, and since that workspace enables flexible integration of everything it contains, it makes sense that its various elements will combine into a unified experience. And on Dehaene's version, ignition into the workspace is a sharp-boundaried event: Information either completely ignites, becoming available for all downstream processes, or it does not. There is no (or only rarely) partial ignition. This can explain determinacy.

But future AI systems might not share this structure. They might have multiple or partially overlapping workspaces. Different specialized subsystems might have access to different regions of a partly-shared workspace. Some animals, such as snails and octopuses, distribute processing among multiple ganglia or neural centers that are less tightly coupled than the hemispheres of the human brain. A robot might broadcast information relevant to locomotion to one area and information relevant to speech to another with limited connectivity.

If the subsystems are entirely disconnected, the result might be entirely discrete centers of subjective experience within a single organism or machine. But if they are partly connected, experience might be only partly unified. In the park bench example, the experience of the trees might be unified with the experience of dampness, and the experience of dampness with memories of yesteryear, but the experience of the trees might not be unified with the memories. (Unification would not then be a transitive relation.) Alternatively, some weaker relation of partial unification might hold among the visual, tactile, and memorial experiences. If this seems inconceivable or impossible, see Sophie Nelson's and my article on indeterminate or fractional subjects.

More abstractly: There's no compelling architectural reason why an AI system would have to make information available either to all downstream processes or to none. A workspace defined in terms of downstream availability could be a patchwork of partial availabilities rather than a fully global all-or-nothing broadcast.

For the same reason, ignition into the workspace needn't be all-or-nothing. Between full ignition with determinate consciousness and no ignition with determinate nonconsciousness, there might be in-between, gray-area half-ignitions that are neither determinately conscious nor determinately nonconscious. Nearly every property with a complex physical or functional basis allows indeterminate, borderline cases: baldness, extraversion, greenness, happiness, whether you're wearing a shoe, whether a country is a democracy. The human global workspace might minimize indeterminacy -- like it's rarely indeterminate in basketball whether the ball has gone through the hoop. But change the architecture and indeterminacy might become common: a half-hearted ignition, or just enough information-sharing to make it indeterminate whether a workspace even exists. (If indeterminacy about consciousness strikes you as inconceivable or impossible, see my 2023 article on borderline consciousness.)

Global Workspace Theory might of course be wrong. But most other theories of consciousness make my argument at least as easy. Dennett's fame-in-the-brain version of broadcast theory explicitly permits disunity and indeterminacy. Higher Order Theories admit the same fragmentation and, probably, gradualism. So do biological theories and theories that focus on embodiment. (Integrated Information Theory is an exception: Its axioms require bright-lined unity and determinacy. But as I've argued, those bright-line axioms lead to unpalatable consequences.)

Recognizing these possibilities for AI systems invites the further thought: Maybe we humans aren't quite as unified as we normally suppose. Maybe indeterminate and disunified consciousness is common. Maybe processes outside of attention hover indeterminately between being conscious and nonconscious. Maybe some processes are only partly unified. If it seems otherwise in introspection and memory, maybe that's because introspection and memory tend to impose unity and determinacy where none was before.

[a Paul Klee painting, untitled 1914: source]

Does Global Workspace Theory Solve the Question of AI Consciousness?

Hint: no.

Below are three sections from Chapter Eight of my manuscript in draft, AI and Consciousness, fresh new version available today here. Comments welcome!

[image adapted from Dehaene et al. 2011]

1. Global Workspace Theories and Access.

The core idea of Global Workspace Theory is simple. Sophisticated cognitive systems like the human mind employ specialized processes that operate to a substantial extent in isolation. We can call these modules, without committing to any strict interpretation of that term.[1] For example, when you hear speech in a familiar language, some cognitive process converts the incoming auditory stimulus into recognizable speech. When you type on a keyboard, motor functions convert your intention to type a word like “consciousness” into nerve signals that guide your fingers. When you try to recall ancient Chinese philosophers, some cognitive process pulls that information from memory without (amazingly) clogging your consciousness with irrelevant information about German philosophers, British prime ministers, rock bands, or dog breeds.

Of course, not all processes are isolated. Some information is widely shared, influencing or available to influence many other processes. Once I recall the name “Zhuangzi”, the thought “Zhuangzi was an ancient Chinese philosopher” cascades downstream. I might say it aloud, type it out, use it as a premise in an inference, form a visual image of Zhuangzi, contemplate his main ideas, attempt to sear it into memory for an exam, or use it as a clue to decipher a handwritten note. To say that some information is in “the global workspace” just is to say that it is available to influence a wide range of cognitive processes. According to Global Workspace Theory, a representation, thought, or cognitive process is conscious if and only if it is in the global workspace – if it is “widely broadcast to other processors in the brain”, allowing integration both in the moment and over time.[2]

Recall the ten possibly essential features of consciousness from Chapter Three: luminosity, subjectivity, unity, access, intentionality, flexible integration, determinacy, wonderfulness, specious presence, and privacy. [Blog readers: You won't have read Chapter Three, but try to ride with it anyway.] Global Workspace Theory treats access as the central essential feature.

Global Workspace theory can potentially explain other possibly essential features. Luminosity follows if processes or representations in the workspace are available for introspective processes of self-report. Unity might follow if there’s only one workspace, so that everything in it is present together. Determinacy might follow if there’s a bright line between being in the workspace and not being in it. Flexible integration might follow if the workspace functions to flexibly combine representations or processes from across the mind. Privacy follows if only you can have direct access to the contents of your workspace. Specious presence might follow if representations or processes generally occupy the workspace for some hundreds of milliseconds.

In ordinary adult humans, typical examples of conscious experience – your visual experience of this text, your emotional experience of fear in a dangerous situation, your silent inner speech, your conscious visual imagery, your felt pains – appear to have the broad cognitive influences Global Workspace Theory describes. It’s not as though we commonly experience pain but find that we can’t report it or act on its basis, or that we experience a visual image of a giraffe but can’t engage in further thinking about the content of that image. Such general facts, plus the theory’s potential to explain features such as luminosity, unity, determinacy, flexible integration, privacy, and specious presence, lend Global Workspace Theories substantial initial attractiveness.

I have treated Global Workspace Theory as if it were a single theory, but it encompasses a family of theories that differ in detail, including “broadcast” and “fame” theories – any theory that treats the broad accessibility of a representation, thought, or process as the central essential feature making it conscious.[3]

Consider two contrasting views: Dehaene’s Global Neuronal Workspace Theory and Daniel Dennett’s “fame in the brain” view. Dehaene holds that entry into the workspace is all-or-nothing. Once a process “ignites” into the workspace, it does so completely. Every representation or process either stops short of entering consciousness or is broadcast to all available downstream processes. Dennett’s fame view, in contrast, admits degrees. Representations or processes might be more or less famous, available to influence some downstream cognitive processes without being available to influence others. There is no one workspace, but a pandemonium of competing processes.[4] If Dennett is correct, luminosity, determinacy, unity, and flexible integration all potentially come under threat in a way they do not as obviously come under threat on Dehaene’s view.[5]

Dennettian concerns notwithstanding, all-or-nothing ignition into a single, unified workspace is currently the dominant version of Global Workspace Theory. The issue remains unsettled and has obvious implications for the types of architectures that might plausibly host AI consciousness.

2. Consciousness Outside the Workspace; Nonconsciousness Within It?

Global Workspace Theory is not the correct theory of consciousness unless all and only thoughts, representations, or processes in the Global Workspace are conscious. Otherwise, something else, or something additional, is necessary for consciousness.

It is not clear that even in ordinary adult humans a process must be in the Global Workspace to be conscious. Consider the case of peripheral experience. Some theorists maintain that people have rich sensory experiences outside of focal attention: a constant background experience of your feet in your shoes and objects in the visual periphery.[6] Others – including Global Workspace theorists – dispute this. Introspective reports vary, and resolving such issues is methodologically tricky.

One methodological problem: People who report constant peripheral experiences might mistakenly assume that such experiences are always present because they are always present whenever they think to check, and the very act of checking might generate those experiences. This is sometimes called the “refrigerator light illusion”, akin to the error of thinking the refrigerator light is always on because it’s always on when you open the door to check.[7] On this view, you’re only tempted to think you have constant tactile experience of your feet in your shoes because you have that experience on those rare occasions when you’re thinking about whether you have it. Even if you now seem to have a broad range of experiences in different sensory modalities simultaneously, this could result from an unusual act of dispersed attention, or from “gist” perception or “ensemble” perception, in which you are conscious of the general gist or general features of a scene, knowing that there are details, without actually experiencing those unattended details.[8]

The opposite mistake is also possible. Those who deny a constant stream of peripheral experiences might simply be failing to notice or remember them. The fact that you don’t remember now the sensation of your feet in your shoes two minutes ago hardly establishes that you lacked the sensation at the time. Although many people find it introspectively compelling that their experience is rich with detail or that it is not, the issue is methodologically complex because introspection and memory are not independent of the phenomena to be observed.[9]

If we do have rich sensory experience outside of attention, it is unlikely that all of that experience is present in or broadcast to a Global Workspace. Unattended peripheral information is rarely remembered or consciously acted upon, tending to exert limited downstream influence – the paradigm of information that is not widely broadcast. Moreover, the Global Workspace is generally characterized as limited capacity, containing only a few thoughts, representations, objects, or processes at a time – those that survive some competition or attentional selection – not a welter of richly detailed experiences in many modalities at once.[10]

A less common but equally important objection runs in the opposite direction: Perhaps not everything in the Global Workspace is conscious. Some thoughts, representations, or processes might be widely broadcast, shaping diverse processes, without ever reaching explicit awareness.[11] Implicit racist assumptions, for example, might influence your mood, actions, facial expressions, and verbal expressions. The goal of impressing your colleagues during a talk might have pervasive downstream effects without occupying your conscious experience moment to moment.

The Global Workspace theorist who wants to allow that such processes are not conscious might suggest that, at least for adult humans, processes in the workspace are generally also available for introspection. But there’s substantial empirical risk in this move. If the correlation between introspective access and availability for other types of downstream cognition isn’t excellent, the Global Workspace theorist faces a dilemma. Either allow many conscious but nonintrospectable processes, violating widespread assumptions about luminosity, or redefine the workspace in terms of introspectability, which amounts to shifting to a Higher Order view.

3. Generalizing Beyond Vertebrates.

The empirical questions are difficult even in ordinary adult humans. But our topic isn’t ordinary adult humans – it’s AI systems. For Global Workspace Theory to deliver the right answers about AI consciousness, it must be a universal theory applicable everywhere, not just a theory of how consciousness works in adult humans, vertebrates, or even all animals.

If there were a sound conceptual argument for Global Workspace Theory, then we could know the theory to be universally true of all conscious entities. Empirical evidence would be unnecessary. It would be as inevitably true as that rectangles have four sides. But as I argued in Chapter Four, conceptual arguments for the essentiality of any of the ten possibly essential features are unlikely to succeed – and a conceptual argument for Global Workspace Theory would be tantamount to a conceptual argument for the essentiality of access, one of those ten features. Not only do the general observations of Chapter Four suggest against a conceptual guarantee, so also does the apparent conceivability, as described in Section 2 above, of consciousness outside the workspace or nonconsciousness within it – even if such claims are empirically false.

If Global Workspace Theory is the correct universal theory of consciousness applying to all possible entities, an empirical argument must establish that fact. But it’s hard to see how such an empirical argument could proceed. We face another version of the Problem of the Narrow Evidence Base. Even if we establish that in ordinary humans, or even in all vertebrates, a thought, representation, or process is conscious if and only if it occupies a Global Workspace, what besides a conceptual argument would justify treating this as a universal truth that holds among all possible conscious systems?

Consider some alternative architectures. The cognitive processes and neural systems of octopuses, for example, are distributed across their bodies, often operating substantially independently rather than reliably converging into a shared center.[12] AI systems certainly can be, indeed often are, similarly decentralized. Imagine coupling such disunity with the capacity for self-report – an animal or AI system with processes that are reportable but poorly integrated with other processes. If we assume Global Workspace Theory at the outset, we can conclude that only sufficiently integrated processes are conscious. But if we don’t assume Global Workspace Theory at the outset, it’s difficult to imagine what near-future evidence could establish that fact beyond a reasonable standard of doubt to a researcher who is initially drawn to a different theory.

If the simplest version of Global Workspace Theory is correct, we can easily create a conscious machine. This is what Dehaene and collaborators envision in the 2017 paper I discussed in Chapter One. Simply create a machine – such as an autonomous vehicle – with several input modules, several output modules, a memory store, and a central hub for access and integration across the modules. Consciousness follows. If this seems doubtful to you, then you cannot straightforwardly accept the simplest version of Global Workspace Theory.[13]

We can apply Global Workspace Theory to settle the question of AI consciousness only if we know the theory to be true either on conceptual grounds or because it is empirically well established as the correct universal theory of consciousness applicable to all types of entity. Despite the substantial appeal of Global Workspace Theory, we cannot know it to be true by either route.

-------------------------------------

[1] Full Fodorian (1983) modularity is not required.

[2] Mashour et al. 2020, p. 776-777.

[3] E.g., Baars 1988; Dennett 1991, 2005; Tye 2000; Prinz 2012; Dehaene 2014; Mashour et al. 2020.

[4] Whether Dennett’s view is more plausible than Dehaene’s turns on whether, or how commonly, representations or processes are partly famous. Some visual illusions, for example, seem to affect verbal report but not grip aperture: We say that X looks smaller than Y, but when we reach toward X and Y we open our fingers to the same extent, accurately reflecting that X and Y are the same size. The fingers sometimes know what the mouth does not. (Aglioti et al. 1995; Smeets et al. 2020). We adjust our posture while walking and standing in response to many sources of information that are not fully reportable, suggesting wide integration but not full accessibility (Peterka 2018; Shanbhag 2023). Swift, skillful activity in sports, in handling tools, and in understanding jokes also appears to require integrating diverse sources of information, which might not be fully integrated or reportable (Christensen et al. 2019; Vauclin et al. 2023; Horgan and Potrč 2010). In response, the all-or-nothing “ignition” view can explain away such cases of seeming intermediacy or disunity as atypical (it needn’t commit to 100% exceptionless ignition with no gray-area cases), by allowing some nonconscious communication among modules (which needn’t be entirely informationally isolated), and/or by allowing for erroneous or incomplete introspective report (maybe some conscious experiences are too brief, complex, or subtle for people to confidently report experiencing them).

[5] Despite developing a theory of consciousness, Dennett (2016) endorsed “illusionism”, which rejects the reality of phenomenal consciousness (see especially Frankish 2016). I interpret the dispute between illusionists and nonillusionists as a verbal dispute about whether the specific philosophical concept of “phenomenal consciousness” requires immateriality, irreducibility, perfect introspectibility, or some other dubious property, or whether the term can be “innocently” used without invoking such dubious properties. See Schwitzgebel 2016, 2025.

[6] Reviewed in Schwitzgebel 2011, ch. 6; and though limited only to stimuli near the center of the visual field, see the large literature on “overflow” in response to Block 2007.

[7] Thomas 1999.

[8] Oliva and Terralba 2006; Whitney and Leib 2018.

[9] Schwitzgebel 2007 explores the methodological challenges in detail.

[10] E.g., Dehaene 2014; Mashour et al. 2020.

[11] E.g., Searle 1983, ch. 5; Bargh and Morsella 2008; Lau 2022; Michel et al. 2025; see also note 4.

[12] Godfrey-Smith 2016; Carls-Diamante 2022.

[13] See also Goldstein and Kirk-Giannini (forthcoming) for an extended application of Global Workspace Theory to AI consciousness. One might alternatively read Dehaene, Lau, and Kouider 2017 purely as a conceptual argument: If all we mean by “conscious” is “accessible in a Global Workspace”, then building a system of this sort suffices for building a conscious entity. The difficulty then arises in moving from that stipulative conceptual claim to the interesting, substantive claim about phenomenal consciousness in the standard sense described in Chapter Two. Similar remarks apply to the Higher Order aspect of that article. One challenge for this deflationary interpretation is that in related works (Dehaene 2014; Lau 2022) the authors treat their accounts as accounts of phenomenal consciousness. The article concludes by emphasizing that in humans “subjective experience coheres with possession” of the functional features they identify. A further complication: Lau later says that the way he expressed his view in this 2017 article was “unsatisfactory”: Lau 2022, p. 168.

Is Signal Strength a Confound in Consciousness Research?

Matthias Michel is among the sharpest critics of the methods of consciousness science. His forthcoming paper, "Consciousness Doesn't Do That", convincingly challenges background assumptions behind recent efforts to discover the causes, correlates, and prevalence of consciousness. It should be required reading for anyone tempted to argue, for example, that trace conditioning correlates with consciousness in humans and thus that nonhuman animals capable of trace conditioning must also be conscious.

But Michel does make one claim that bugs me, and that claim is central to the article. And Hakwan Lau -- another otherwise terrific methodologist -- makes a similar claim in his 2022 book In Consciousness We Trust, and again the claim is central to the argument of that book. So today I'm going to poke at that claim, and maybe it will burst like a sour blueberry.

The claim: Signal strength (performance capacity, in Lau's version) is a confound in consciousness research.

As Michel uses the phrase, "signal strength" is how discriminable a perceptible feature is to a subject. A sudden, loud blast of noise has high signal strength. It's very easy to notice. A faint wavy pattern in a gray field, presented for a tenth of second, has low signal strength. It is easy to miss. Importantly, signal strength is not the same as (objective, externally measurable) stimulus intensity, but reflects how well the perceiver responds to the signal.

Signal strength clearly correlates with consciousness. You're much more likely to be conscious of stimuli that you find easy to discriminate than stimuli that you find difficult to discriminate. The loud blare is consciously experienced. The faint wavy pattern might or might not be. A stimulus with effectively zero signal strength -- say, a gray dot flashed for a millionth of a second and immediately masked -- will normally not be experienced at all.

But signal strength is not the same as consciousness. The two can come apart. The classic example is blindsight. On the standard interpretation (but see Phillips 2020 for an alternative), patients with a specific type of visual cortex damage can discriminate stimuli that they cannot consciously perceive. Flash either an "X" or an "O" in the blind part of their visual field and they will say they have no visual experience of it. But ask them to guess which letter was shown and their performance is well above chance -- up to 90% correct in some tasks. The "X" has some signal strength for them: It's discriminable but not consciously experienced.

If signal strength is not consciousness but often correlates with it, the following worry arises. When a researcher claims that "trace conditioning is only possible for conscious stimuli" or "consciousness facilitates episodic memory", how do you know that it's really consciousness doing the work, rather than signal strength? Maybe stimuli with high signal strength are both more likely to be consciously experienced and more likely to enable trace conditioning and episodic memory. Unless researchers have carefully separated the two, the causal role of consciousness remains unclear.

An understandable methodological response is to try to control for signal strength: Present stimuli of similar discriminability to the subject but which differ in whether (or to what extent) they are consciously experienced. Only then, the reasoning goes, can differences in downstream effects be confidently attributed to consciousness itself rather than differences in signal strength. Lau in particular stresses the importance of such controls. Yet such careful matching is difficult and rarely attempted. On this reasoning, much of the literature on the cognitive role of consciousness is built on sand, not clearly distinguishing the effects of consciousness from the effects of signal strength.

This reasoning is attractive but faces an obvious objection, which both Michel and Lau address directly. What if signal strength just is consciousness? Then trying to "control" for it would erase the phenomenon of interest.

Both Michel and Lau analogize to height and bone length. If you want to test whether height confers an advantage in basketball or dating, you might want to control for skin color, but it would be absurd to control for bone length. If skin color correlates with height and you want to see whether height specifically advantages people in basketball or dating, it makes sense to control for differences in skin color by systematically comparing people with the same skin color but different heights. If the advantage persists, you can infer that height rather than skin color is doing the work. But trying to control for bone length lands you in nonsense. Taller people just are the people with longer bones.

Michel and Lau respond by noting that consciousness and signal strength (or performance capacity) sometimes dissociate, as in blindsight. Therefore, they are not the same thing and it does make sense to control for one in exploring the effects of the other.

But this response is too simple and too fast.

We can see this even in their chosen example. Height and bone length aren't quite the same thing. They can dissociate. People are about 1-2 cm taller in the morning than at night -- not because their bones have grown but because the tissue between the bones (especially in the spine) compresses during the day.

Now imagine an argument parallel to Michel's and Lau's: Since height and bone length can come apart, we should try to control for bone length in examining the effects of height on basketball and dating. We then compare the same people's basketball and dating outcomes in the morning and at night, "holding bone length fixed" while height varies slightly. This would be a methodological mistake. For one thing, we've introduced a new potential confound, time of day. For another, even if the centimeter in the morning really does help a little, we've dramatically reduced our ability to detect the real effect of height by "overcontrolling" for a component of the target variable, height.

Consider a psychological example. The personality trait of extraversion can be broken into "facets", such as sociability, assertiveness, and energy level. Since energy level is only one aspect of extraversion, the two can dissociate. Some people are energetic but not sociable or assertive; others are sociable and assertive but low-energy. If you wanted to measure the influence of extraversion on, say, judgments of likeability in the workplace, you wouldn't want to control for energy level. That would be overcontrol, like controlling for bone length in attempting to assess the effects of height. It would strip away part of the construct you are trying to measure.

What I hope these examples make clear is that dissociability between correlates A and B does not automatically make B a confound that must be controlled when studying A's effects. Bone length is dissociable from height, but it is a component, not a confound. Energy level is dissociable from extraversion, but it is a component, not a confound.

The real question, then, is whether signal strength (or performance capacity) is better viewed as a component or facet of consciousness than as a separate variable that needs to be held constant in testing the effects of consciousness.

A case can be made that it is. Consider Global Workspace Theory, one of the leading theories of consciousness. On this view, a process or representation is conscious if it is broadly available for "downstream cognition" such as verbal report, long-term memory, and rational planning. If discrimination judgments are among those downstream capacities, then one facet of being in the global workspace (that is, on this view, being conscious) is enabling such judgments. But recall that signal strength just is discriminability for a subject. If so, things begin to look like the extraversion / energy case. Controlling for discriminability would be overcontrolling, that is, attempting to equalize or cancel the effects not of a separate, confounding process, but of a component of the target process itself. (Similar remarks hold for Lau's "performance capacity".)

Global Workspace Theory might not be correct. And if it's not, maybe signal strength is indeed a confounder, rather than a component of consciousness. But the case for treating signal strength as a confounder can't be established simply by noticing the possibility of dissociations between consciousness and signal strength. Furthermore, since Michel's and Lau's recommended methodology can be trusted not to suffer from overcontrol bias only if Global Workspace Theory is false, it's circular to rely on that methodology to argue against Global Workspace Theory.